A Software Perspective

Green IT A Software Perspective

Victor Bahl

SIGCOMM Green Workshop, August 30, 2010

Acknowledgments & Credit

Goraczko Agarwal Chandra

Nath Liu Kansal

Padhye Reich Zhao

Victor Bahl, SIGCOMM Green Networking2 Workshop 2010 Environmental Sustainability

REDUCING MANAGING RETHINKING energy demands energy and business environmental practices footprint

…while improving the bottom line

Victor Bahl, SIGCOMM Green Networking3 Workshop 2010 Home & Office Computing

Battery Powered Computers “Wall Powered” Computers

Lenovo X61 laptop Dell Optiplex 745 desktop Power: 0.74W (sleep) to 16W (active) Power: 1.2W (sleep) to >140W (active)

• 67% of office PCs are left on after work hours – Robertson et. al.: After-hour power status of office equipment and energy usage of plug-load devices. LBNL report #53729 • Home PCs are left on for 34% of the time ̵ 50% of the time they are not being used • 600+ desktops always left on (total=700+ ) – Agarwal et. al: Sombiloqui, Augmenting network Interfaces to reduce PC energy (NSDI 2009) • Almost all desktop machines in MSR are left on after work hours Victor Bahl, SIGCOMM Green Networking Workshop 2010 Enterprise & Datacenter Computing

Source: US Dept. of Energy

40 30 Est. Cost: 20 $4.5B

10 Energy usage

growing at 14%

Electricity (billion kWh/yr) Electricity

yearly

2000 2003 2006

Datacenter energy (excluding small DC’s, office IT equip.) equals electricity used by the entire U.S transportation manufacturing industry (manufacture of automobiles, aircraft, trucks, and ships)

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Long Term Energy Forecasts

• EPA estimated, in 2006 datacenters in the US used 61 billion KWH or 1.5% of the energy used in the country

• EPA’s Prediction: By 2011, datacenters in the US would use up to 100 billion KWH with a total cost of $7.4 billion and an estimated emissions impact of 59 million metric tons of CO2

• By 2030 214 GigaWatts of new generation capacity are expected to be needed

• The electric utility industry will need to make a total infrastructure investment of $1.5 - 2.0 Trillion

Energy efficiency/demand response programs by electric utilities could reduce the need for new generation capacity significantly; dropping the forecast by 38 percent Source: Energy Information Administration, EPRI

Victor Bahl, SIGCOMM Green Networking6 Workshop 2010 Opportunities to Reduce Power A Holistic View

Improve Data Center PUE Building Increase Facility Utilization

Management Increase Server Utilization Infrastructure Centralized Power Management

Hardware Rightsizing & Subsystem Balancing Package Power Supply Efficiency Remove Unnecessary Components Applications Energy Efficient & Aware Applications

Operating System Platform Power Management Energy Source Aware

Silicon Variable-power components Lower-power components

PUE: Power Usage Effectiveness (ratio of total power over IT power) Victor Bahl, SIGCOMM Green Networking Workshop 2010 In the beginning….

SmartPhone Energy Management

….was all about sleeping Wake on Wireless (2001-02) LPR to a Trigger Wakeups

Switch On/Off iPAQ power monitor

Power GND

iPAQ Rx

Accelerometer 915 MHz Radio Crystal Tilt Sensor IR Range Speaker

PIC

Temperature Vibrator Sensor

Audio Plug Front View Back View Victor Bahl, SIGCOMM Green Networking Workshop 2010 MobiCom 2002 Wake on Wireless with Low Power Radio Energy Saving

2001-02

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Wake on Wireless (2005-07) MobiSys 2007 Cell Network as a Trigger Mechanism

– Turn off Wi-Fi IP Phone Soft Phone interface if Enterprise Network Smartphone not in Access VoIP call Point Internet Smart Phone LAN – Notify incoming call ATA SIP Server using “ring” on GSM interface PSTN • Turn on Wi-Fi interface and complete call over Wi-Fi

GSM Network – Turn off Wi-Fi Wi -Fi interface GSM interface interface on call Base Station completion

Security: SIP Server uses different caller ID every time

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Wake on Wireless with Cell2Notify MobiSys 2007 Energy Savings

60 60 50 John 50 James 40 40 30 30 20 20 10 10

Duration of CallsofDuration (Minutes) 0 Duration of CallsofDuration (Minutes) 0 1 3 5 7 9 11 13 15 17 19 21 23 1 3 5 7 9 11 13 15 17 19 21 23 Hour of the Day Hour of the Day 70 60 Using WiFi Using Cell2Notify 50 40 30 20

10 Lifetime (HoursLifetime Usage)of 0 John James Victor Bahl, SIGCOMM Green Networking Workshop 2010 … then we extended sleeping to home & enterprise

Home & Enterprise Energy Management

• WaMu, Dell and GE save of millions of dollars per year w./ propriety solutions (Citation: Dog Washburn: How much money are your idle PC wasting (Forrester, December 2008)

• UCSD CS department estimates it can cut 25% of total energy bill (Agarwal et. al: Somniloqui, Augmenting network Interfaces to reduce PC energy usage (NSDI 2008)

Victor Bahl, SIGCOMM Green Networking Workshop 2010 NSDI 2008 Somniloquy USB Interface (Wake up Host + Status + Debug)

Uses “gumstix” platform USB Interface – PXA270 processor with full TCP/IP stack (power + USBNet) – USB connection to PC for sleep detection/wakeup trigger, power while asleep, and IP networking for data

SD Storage

Processor

100Mbps Ethernet Interface

Victor Bahl, SIGCOMM Green Networking Workshop 2010 NSDI 2008 Somniloquy: PCs that Talk in their Sleep

• Augment network interfaces: – Add a separate power domain • Powered on when host is asleep Host PC Somniloquy Apps • Processor + Memory + daemon Flash Storage + Network stack Operating system, including networking – Same MAC/IP Address stack wakeup Appln. Host processor, filters stubs RAM, peripherals, etc. • Wake up Host when needed Embedded OS, including – E.g. incoming connection Secondary processor networking stack Embedded Network interface CPU, RAM, hardware flash • Handle some applications while PC remains asleep

– Using “application stubs” Victor Bahl, SIGCOMM Green Networking Workshop 2010 Somniloquy Power Savings

State Power Normal Idle State 102.1W Lowest CPU frequency 97.4W Disable Multiple cores 93.1W “Base Power” 93.1W Suspend state (S3) 1.2W

For Desktops – Power drops from >100W to <5W Dell Optiplex 745 Power Consumption and transitions between states – Assuming a 45 hour work week . 620kWh saved per year

. US $56 savings, 378 kg CO2

For Laptops – Power drops from >11W to 1W, . Battery life increases from <6 hours to >60 hours – Provides functionality of the “Baseline” state – Power consumption similar to “Sleep” state Victor Bahl, SIGCOMM Green Networking Workshop 2010 Enterprise IT Dominates Energy Usage

UCSD CSE Building Microsoft Research, Bldg. 99

Courtesy: Yuvraj Agarwal (UCSD)

• IT loads account for 50% (peak) - 80% (off-peak) loads! – Includes Plug-Loads (PCs) + Machine Room (Servers)

• Most IT equipment not powered off or `Sleeping’

Victor Bahl, SIGCOMM Green Networking Workshop 2010 SleepProxy USENIX ATC 2010 A Software-only Enterprise Solution Subnet router

WAN

TCP SYN 1.2.3.4:3389TCP SYN 1.2.3.4:3389

WOL / Magic Packet 00:11:22:33:44:55 … Sleep notification 00:11:22:33:44:55 1.2.3.4 Listing ports: 445, 3389 ARP Probe 00:11:22:33:44:55 1.2.3.4 ARP Probe SYN-ACK00:11:22:33:44:55 Sleep Server 1.2.3.4 Typical machine sleeps > 40% of the time LiteGreen USENIX ATC 2010 A VM Based Solution

Live Migration RDP client VM VM Virtualization Server Hardware

• Prevalence of short idle periods • Virtualization + Migration to save energy during short and long idle periods while avoiding user disruption • LiteGreen can help save 60-72% of desktop energy

Victor Bahl, SIGCOMM Green Networking Workshop 2010 You Can’t Manage what you don’t Measure

Turning our attention to transparency & possibly changing behavior….

Streamline data collection for key environmental performance indicators:  Direct energy consumption  Indirect energy consumption  Greenhouse gas emissions from the total energy consumption  Greenhouse gas emissions from standard business practices Analyze status of environmental sustainability initiatives Share performance broadly across the organization

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Microsoft Hohm Gartner Inc. survey: 80% of consumers (in US & U.K) would like to participate in some energy program if it were offered by providers (Nay 2009)

Victor Bahl, SIGCOMM Green Networking Workshop 2010

Fine Grained Visibility Measure application energy usage in depth

Detailed Energy Run-time control Profile Source code (Eg. Reduce QoS)

JouleMeter Auto-optimize Workload (Eg. Compiler warning: Change line 136 in Prog.c Designer Insights to save …) (Eg. Network is holding up CPU in method X …) Optimizations

GreenClient: Provides energy usage stats to user - Current power draw - Historical savings

Courtesy: Aman Kansal SOCC 2010 Providing Transparency with Joulemeter Application Energy Measurement Using Performance Events Power Meter Color label Issues Disk B write A O A A B A CPU S

Memory A B B A A A B

OS commits buffered Disk writes to idle disk Spin Up A B

• Energy(App A) = S (red rectangles) – Other components: Network, GPU, … (Not in current version) Victor Bahl, SIGCOMM Green Networking Workshop 2010 GreenUP: JouleMeter + SleepServer

Victor Bahl, SIGCOMM Green Networking Workshop 2010 GreenUP: SleepServer + JouleMeter(2)

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Elsewhere

Courtesy: Yuvraj Agarwal (UCSD) Victor Bahl, SIGCOMM Green Networking Workshop 2010 ……….taking on the big beast

Datacenter Energy Management

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Rethinking Business Practices Reduce Travel, Commuting, and Office Space Requirements

Reduce datacenter workload/energy consumption with cloud services Maximize efficiency and sustainability with datacenter operation efficiency

Transfer some of the environmental footprint to Advantage of scale — odds are DCs used a third party data center to run your workloads are more efficient than equipment inyour own premises.

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Efficiencies of Scale

Microsoft’s Quincy, WA, data center uses 100 percent renewable hydropower. Microsoft’s San Antonio, TX, data center uses recycled water for cooling & wind as a primary energy source.

Dublin Data Center

1.25 PUE • First mega data center outside US (303K sq. ft & growing) • European Commission Best Practice Award • Free air-cooling - Chiller-free, air-side economization - 1% water use - Saves 18 million liters of water/month - 50% more efficient 5.4 MW => 22.2 MW • Reduced waste and carbon Footprint

Chicago Data Center

1.22 PUE • Largest known container data center (700,000 sq. ft) • Scalable & modular, reduced capital costs, construction, packaging, and energy waste • Water side economizers • Reduced waste and lower carbon footprint 30 MW => 60 MW Victor Bahl, SIGCOMM Green Networking Workshop 2010 200 + Sites and Services DC Operation Efficiency

In poorly designed DC, 50% of the energy goes into Air-conditioning

30% of the servers can be turned off for Messenger traffics (NSDI Courtesy: James Hamilton (Amazon) 2008)

Relative Data Center Energy Losses Relative PC Energy Losses 0% 0% 3% Building 5% 3% 13% 11% Management Infra Management Infra 15% 30% Chassis 6% Server Chassis 13% Applications Applications 11% 21% OS OS Silicon Silicon

Source: Microsoft estimates Victor Bahl, SIGCOMM Green Networking Workshop 2010 DC Genome System Courtesy: Jie Liu

Real-time Monitoring Facility layout & Control DC Genome Data Explorer Change Archiving Cooling Systems Management

Power Systems Capacity Data Planning staging “Genome” Dynamic Heat Distribution (Classes, labels, relations, dynamics, Server Data analysis & probabilistic models) Provisioning Server Utilizations Optimization Tools Fault

Data Center Diagnostics/ Software Tolerance Infrastructure

Network Activities Data Visualization SenSys 2009

Computer Room AC – 4 Temperature Contour Maps Front

Back Psychrometric charts

95F

Dell PowerEdge 2950 Server Warranty

81F

ASHRAE 2008 Recommended Class 1

ASHRAE: American Society of Heating, Refrigerating and Air-Conditioning Engineers Efficiency vs. Density

Greater Server Density = Increased Energy/Ft3 Fewer More Watts Trans- per actions Trans- per action Watt Less Power Per Server Victor Bahl, SIGCOMM Green Networking Workshop 2010 Reducing Energy Demands Utilization/Reduce the Number of Physical Machines

Gain multiple virtual servers for the energy/power cost of one physical server

12% utilization

11% utilization 6% utilization

18% utilization 80% utilization 13% utilization 5% utilization

3% utilization

2% utilization 3% utilization

7% utilization

“...the effective use of virtualization can reduce server energy consumption by as much as 82 percent and floor space by as much as 86 percent.” Gartner, Inc., June 2009 Victor Bahl, SIGCOMM Green Networking Workshop 2010 Sample Savings from Virtualization

18 Online Cloud Online 100 16 14 50 12

10 (%) CPU 0 8 Traces from MSN Messenger 6 Savings (%) Savings 4

2 Cloud-like

0 10 99 95 90 80

Percentile Provisioned 5 CPU (%) CPU Reclaimed Savings 0 Time(s) 13% to 27%, with benefits of virtualization Traces from AdCenter, Creative Asset Management , MSN TV Apollo , Ad Note: Virtualization is a networking problem Delivery , Shared Services group

Victor Bahl, SIGCOMM Green Networking Workshop 2010 …..moving beyond sleeping

…and beyond Data centers Rethinking Business Practices Reduce Travel, Commuting, and Office Space Requirements

Travel & Commuting Replace travel and • 62% of the companies are commuting for meetings already reducing their business or training with Web/ travel footprint + 24% are videoconferencing developing plans to do so. • 89% expect they will want to fly less over the next 10 years. Support tele- and remote workers with portals, Travelling Light, U.K. WWF, social computing, and 2008 document workspaces

“... if a significant number of people worked from home more than three days a week, this could lead to energy savings of 20-50%...” The Climate Group, 2008

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Telepresence A videoconferencing experience that creates the illusion that the remote participants are in the same room with you.

Videoconferencing has the potential to replace up to 30 percent of business travel The Climate Group (SMART 2020, June 2008)

Cisco Telepresence $299k Tandberg Experia $225k

HP Halo $425k + $18k/mo Polycom RPX210M $269k + $18.5k/mo Victor Bahl, SIGCOMM Green Networking Workshop 2010 MS Personal Telepresence Station

Replicate experience people enjoy in face-to-face meeting • Dedicated camera-display-speaker for remote stand-in • Tailored to a single person  Correct spatial cues

Prototypes Full Mesh Geometry

video

Courtesy: Zhengyou Zhang Victor Bahl, SIGCOMM Green Networking Workshop 2010 IT for Greening

Traffic congestion  time & fuel waste (Courtesy: Mario Gerla) . Air pollution caused by vehicles reduces city livability . Existing traffic control through “green wave” is not enough . Nitric Oxide (NO) concentration exhibits high spatial and temporal variability (Environmental Health Perspectives, Nov 2009) o How to implement dynamic congestion and pollution control?

Victor Bahl, SIGCOMM Green Networking Workshop 2010 In collaboration with Greening with TV Whitespaces

~4x range compared to 2.4 GHz (Wi-Fi) with same transmit power and receiver sensitivity

MSR’s Redmond Campus Raw received power at different Route taken by the shuttle (0.95 miles x 0.75 miles) Distances from the transmitter

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Total power consumption of DoCoMo’s Network is 2.4 TWh, that is 5%of Total IT Energy Consumption in Japan. Approx. 126 Wh/day/docomo-user - Atsushi Murase, Managing Director NTT DoCoMo.

Mobile Networks occupy 10% of total IT Consumption & 0.1% of total energy consumption in Japan. - Study Group on ICT , March 2007

Let’s talk about marrying the mobile user to the cloud?

http://research.microsoft.com/hawaii Final Remarks • Look at the problem holistically, lots of opportunities availability – Silicon, OS, applications, hardware packaging, buildings etc. • Greatest savings come from sleeping (works for cell phones, homes, enterprises, datacenters) – My mother taught me “Don’t Waste” -> applies to energy too! • Faster adoption when technology saves significant money – Users adopt it if it doesn’t change their lifestyle . People want to be able to access their machines whenever they want without modifying their own behavior • Transparency can change habits & is necessary for improving architecture & design • Offloading computations to the cloud by SmartPhones has energy implications that still need to be quantified

Victor Bahl, SIGCOMM Green Networking Workshop 2010

It’s the Right Thing to do & It’s Good Business

“Green IT has reached critical mass. Virtually “Enterprise PCs are wasting money. Far all the companies we surveyed too many organizations leave economic and (97 percent) are discussing their environmental value on the table by not Green strategy.” reducing PC-related “Green IT Report Regional Data – energy costs.” United States and Canada: “How Much Money Are Your Idle PCs Survey Results,” Wasting?,” Forrester Research, Symantec, May 2009 Inc., December 2008

“Transformation in the “IT's role will increasingly be way people and businesses about applying the technology use technology could reduce to create more-energy- annual man-made global efficient, less-carbon- emissions by 15 per cent by 2020 intensive business models, and deliver energy efficiency savings to global enterprises, value chains, products and businesses of over EUR 500 billion services with reduced environmental impact.” [GBP 400 billion/USD 800 billion].” “User Survey Analysis: Sustainability and Green IT, Worldwide, 2009,” Gartner, Inc., April 2009 “SMART 2020: Enabling the Low Carbon Economy in the Information Age,” The Climate Group, June 2008

Victor Bahl, SIGCOMM Green Networking Workshop 2010 Some MSR supported Projects with Academia http://www.microsoft.com/environment/research/ Thanks

http://research.microsoft.com/nrg/

© 2007 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.